🐛 | Fix loading wrong model.

AudioUtils.py

@@ -16,3 +16,56 @@ def stretch_tensor(tensor, target_length):
     tensor = F.interpolate(tensor, scale_factor=scale_factor, mode='linear', align_corners=False)
 
     return tensor
+
+def pad_tensor(audio_tensor: torch.Tensor, target_length: int = 128):
+    current_length = audio_tensor.shape[-1]
+
+    if current_length < target_length:
+        padding_needed = target_length - current_length
+
+        padding_tuple = (0, padding_needed)
+        padded_audio_tensor = F.pad(audio_tensor, padding_tuple, mode='constant', value=0)
+    else:
+        padded_audio_tensor = audio_tensor
+
+    return padded_audio_tensor
+
+def split_audio(audio_tensor: torch.Tensor, chunk_size: int = 128) -> list[torch.Tensor]:
+    if not isinstance(chunk_size, int) or chunk_size <= 0:
+        raise ValueError("chunk_size must be a positive integer.")
+
+    # Handle scalar tensor edge case if necessary
+    if audio_tensor.dim() == 0:
+        return [audio_tensor] if audio_tensor.numel() > 0 else []
+
+    # Identify the dimension to split (usually the last one, representing time/samples)
+    split_dim = -1
+    num_samples = audio_tensor.shape[split_dim]
+
+    if num_samples == 0:
+        return [] # Return empty list if the dimension to split is empty
+
+    # Use torch.split to divide the tensor into chunks
+    # It handles the last chunk being potentially smaller automatically.
+    chunks = list(torch.split(audio_tensor, chunk_size, dim=split_dim))
+
+    return chunks
+
+def reconstruct_audio(chunks: list[torch.Tensor]) -> torch.Tensor:
+    if not chunks:
+        return torch.empty(0)
+
+    if len(chunks) == 1 and chunks[0].dim() == 0:
+        return chunks[0]
+
+    concat_dim = -1
+
+    try:
+        reconstructed_tensor = torch.cat(chunks, dim=concat_dim)
+    except RuntimeError as e:
+        raise RuntimeError(
+            f"Failed to concatenate audio chunks. Ensure chunks have compatible shapes "
+            f"for concatenation along dimension {concat_dim}. Original error: {e}"
+        )
+
+    return reconstructed_tensor

README.md

@@ -18,6 +18,7 @@ SISU (Super Ingenious Sound Upscaler) is a project that uses GANs (Generative Ad
 1. **Set Up**:
    - Make sure you have Python installed (version 3.8 or higher).
    - Install needed packages: `pip install -r requirements.txt`
+   - Install current version of PyTorch (CUDA/ROCm/What ever your device supports)
 
 2. **Prepare Audio Data**:
    - Put your audio files in the `dataset/good` folder.

app.py

@@ -0,0 +1,60 @@
+import torch
+import torch.nn as nn
+import torch.optim as optim
+
+import torch.nn.functional as F
+import torchaudio
+import tqdm
+
+import argparse
+import math
+import os
+
+import AudioUtils
+from generator import SISUGenerator
+
+
+# Init script argument parser
+parser = argparse.ArgumentParser(description="Training script")
+parser.add_argument("--device", type=str, default="cpu",  help="Select device")
+parser.add_argument("--model", type=str, help="Model to use for upscaling")
+parser.add_argument("--clip_length", type=int, default=1024, help="Internal clip length, leave unspecified if unsure")
+parser.add_argument("-i", "--input", type=str, help="Input audio file")
+parser.add_argument("-o", "--output", type=str, help="Output audio file")
+
+args = parser.parse_args()
+
+device = torch.device(args.device if torch.cuda.is_available() else "cpu")
+print(f"Using device: {device}")
+
+generator = SISUGenerator()
+
+models_dir = args.model
+clip_length = args.clip_length
+input_audio = args.input
+output_audio = args.output
+
+if models_dir:
+    generator.load_state_dict(torch.load(f"{models_dir}", map_location=device, weights_only=True))
+else:
+    print(f"Generator model (--model) isn't specified. Do you have the trained model? If not you need to train it OR acquire it from somewhere (DON'T ASK ME, YET!)")
+
+generator = generator.to(device)
+
+def start():
+    # To Mono!
+    audio, original_sample_rate = torchaudio.load(input_audio, normalize=True)
+    audio = AudioUtils.stereo_tensor_to_mono(audio)
+
+    splitted_audio = AudioUtils.split_audio(audio, clip_length)
+    splitted_audio_on_device = [t.to(device) for t in splitted_audio]
+    processed_audio = []
+
+    for clip in tqdm.tqdm(splitted_audio_on_device, desc="Processing..."):
+        processed_audio.append(generator(clip))
+
+    reconstructed_audio = AudioUtils.reconstruct_audio(processed_audio)
+    print(f"Saving {output_audio}!")
+    torchaudio.save(output_audio, reconstructed_audio.cpu().detach(), original_sample_rate)
+
+start()

data.py

@@ -5,49 +5,42 @@ import torchaudio
 import os
 import random
 import torchaudio.transforms as T
+import tqdm
 import AudioUtils
 
 class AudioDataset(Dataset):
     audio_sample_rates = [11025]
-    MAX_LENGTH = 44100  # Define your desired maximum length here
 
-    def __init__(self, input_dir, device):
-        self.input_files = [os.path.join(root, f) for root, _, files in os.walk(input_dir) for f in files if f.endswith('.wav')]
+    def __init__(self, input_dir, device, clip_length = 1024):
         self.device = device
+        input_files = [os.path.join(root, f) for root, _, files in os.walk(input_dir) for f in files if f.endswith('.wav') or f.endswith('.mp3') or f.endswith('.flac')]
 
-    def __len__(self):
-        return len(self.input_files)
-
-    def __getitem__(self, idx):
-        # Load high-quality audio
-        high_quality_audio, original_sample_rate = torchaudio.load(self.input_files[idx], normalize=True)
+        data = []
+        for audio_clip in tqdm.tqdm(input_files, desc=f"Processing {len(input_files)} audio file(s)"):
+            audio, original_sample_rate = torchaudio.load(audio_clip, normalize=True)
+            audio = AudioUtils.stereo_tensor_to_mono(audio)
 
             # Generate low-quality audio with random downsampling
             mangled_sample_rate = random.choice(self.audio_sample_rates)
             resample_transform_low = torchaudio.transforms.Resample(original_sample_rate, mangled_sample_rate)
-        low_quality_audio = resample_transform_low(high_quality_audio)
-
             resample_transform_high = torchaudio.transforms.Resample(mangled_sample_rate, original_sample_rate)
-        low_quality_audio = resample_transform_high(low_quality_audio)
 
-        high_quality_audio = AudioUtils.stereo_tensor_to_mono(high_quality_audio)
-        low_quality_audio = AudioUtils.stereo_tensor_to_mono(low_quality_audio)
+            low_audio = resample_transform_low(audio)
+            low_audio = resample_transform_high(low_audio)
 
-        # Pad or truncate high-quality audio
-        if high_quality_audio.shape[1] < self.MAX_LENGTH:
-            padding = self.MAX_LENGTH - high_quality_audio.shape[1]
-            high_quality_audio = F.pad(high_quality_audio, (0, padding))
-        elif high_quality_audio.shape[1] > self.MAX_LENGTH:
-            high_quality_audio = high_quality_audio[:, :self.MAX_LENGTH]
+            splitted_high_quality_audio = AudioUtils.split_audio(audio, clip_length)
+            splitted_high_quality_audio[-1] = AudioUtils.pad_tensor(splitted_high_quality_audio[-1], clip_length)
 
-        # Pad or truncate low-quality audio
-        if low_quality_audio.shape[1] < self.MAX_LENGTH:
-            padding = self.MAX_LENGTH - low_quality_audio.shape[1]
-            low_quality_audio = F.pad(low_quality_audio, (0, padding))
-        elif low_quality_audio.shape[1] > self.MAX_LENGTH:
-            low_quality_audio = low_quality_audio[:, :self.MAX_LENGTH]
+            splitted_low_quality_audio = AudioUtils.split_audio(low_audio, clip_length)
+            splitted_low_quality_audio[-1] = AudioUtils.pad_tensor(splitted_low_quality_audio[-1], clip_length)
 
-        high_quality_audio = high_quality_audio.to(self.device)
-        low_quality_audio = low_quality_audio.to(self.device)
+            for high_quality_sample, low_quality_sample in zip(splitted_high_quality_audio, splitted_low_quality_audio):
+                data.append(((high_quality_sample, low_quality_sample), (original_sample_rate, mangled_sample_rate)))
 
-        return (high_quality_audio, original_sample_rate), (low_quality_audio, mangled_sample_rate)
+        self.audio_data = data
+
+    def __len__(self):
+        return len(self.audio_data)
+
+    def __getitem__(self, idx):
+        return self.audio_data[idx]

discriminator.py

@@ -2,23 +2,34 @@ import torch
 import torch.nn as nn
 import torch.nn.utils as utils
 
-def discriminator_block(in_channels, out_channels, kernel_size=3, stride=1, dilation=1, spectral_norm=True):
+def discriminator_block(in_channels, out_channels, kernel_size=3, stride=1, dilation=1, spectral_norm=True, use_instance_norm=True):
     padding = (kernel_size // 2) * dilation
-    conv_layer = nn.Conv1d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, dilation=dilation, padding=padding)
+    conv_layer = nn.Conv1d(
+        in_channels,
+        out_channels,
+        kernel_size=kernel_size,
+        stride=stride,
+        dilation=dilation,
+        padding=padding
+    )
+
     if spectral_norm:
         conv_layer = utils.spectral_norm(conv_layer)
-    return nn.Sequential(
-        conv_layer,
-        nn.LeakyReLU(0.2, inplace=True),
-        nn.BatchNorm1d(out_channels)
-    )
+
+    layers = [conv_layer]
+    layers.append(nn.LeakyReLU(0.2, inplace=True))
+
+    if use_instance_norm:
+        layers.append(nn.InstanceNorm1d(out_channels))
+
+    return nn.Sequential(*layers)
 
 class AttentionBlock(nn.Module):
     def __init__(self, channels):
         super(AttentionBlock, self).__init__()
         self.attention = nn.Sequential(
             nn.Conv1d(channels, channels // 4, kernel_size=1),
-            nn.ReLU(),
+            nn.ReLU(inplace=True),
             nn.Conv1d(channels // 4, channels, kernel_size=1),
             nn.Sigmoid()
         )
@@ -28,31 +39,25 @@ class AttentionBlock(nn.Module):
         return x * attention_weights
 
 class SISUDiscriminator(nn.Module):
-    def __init__(self, layers=4): #Increased base layer count
+    def __init__(self, base_channels=16):
         super(SISUDiscriminator, self).__init__()
+        layers = base_channels
         self.model = nn.Sequential(
-            discriminator_block(1, layers, kernel_size=3, stride=1), #Aggressive downsampling
-            discriminator_block(layers, layers * 2, kernel_size=5, stride=2),
-            discriminator_block(layers * 2, layers * 4, kernel_size=5, dilation=4),
-
-            #AttentionBlock(layers * 4), #Added attention
-
-            #discriminator_block(layers * 4, layers * 8, kernel_size=5, dilation=4),
-            #AttentionBlock(layers * 8), #Added attention
-            #discriminator_block(layers * 8, layers * 16, kernel_size=5, dilation=8),
-            #discriminator_block(layers * 16, layers * 16, kernel_size=3, dilation=1),
-            #discriminator_block(layers * 16, layers * 8, kernel_size=3, dilation=2),
-            #discriminator_block(layers * 8, layers * 4, kernel_size=3, dilation=1),
-            discriminator_block(layers * 4, layers * 2, kernel_size=5, stride=2),
-            discriminator_block(layers * 2, layers, kernel_size=3, stride=1),
-            discriminator_block(layers, 1, kernel_size=3, stride=1, spectral_norm=False) #last layer no spectral norm.
+            discriminator_block(1, layers, kernel_size=7, stride=1, spectral_norm=True, use_instance_norm=False),
+            discriminator_block(layers, layers * 2, kernel_size=5, stride=2, spectral_norm=True, use_instance_norm=True),
+            discriminator_block(layers * 2, layers * 4, kernel_size=5, stride=1, dilation=2, spectral_norm=True, use_instance_norm=True),
+            AttentionBlock(layers * 4),
+            discriminator_block(layers * 4, layers * 8, kernel_size=5, stride=1, dilation=4, spectral_norm=True, use_instance_norm=True),
+            discriminator_block(layers * 8, layers * 4, kernel_size=5, stride=2, spectral_norm=True, use_instance_norm=True),
+            discriminator_block(layers * 4, layers * 2, kernel_size=3, stride=1, spectral_norm=True, use_instance_norm=True),
+            discriminator_block(layers * 2, layers, kernel_size=3, stride=1, spectral_norm=True, use_instance_norm=True),
+            discriminator_block(layers, 1, kernel_size=3, stride=1, spectral_norm=False, use_instance_norm=False)
         )
+
         self.global_avg_pool = nn.AdaptiveAvgPool1d(1)
-        self.sigmoid = nn.Sigmoid()
 
     def forward(self, x):
         x = self.model(x)
         x = self.global_avg_pool(x)
-        x = x.view(-1, 1)
-        x = self.sigmoid(x)
+        x = x.view(x.size(0), -1)
         return x

file_utils.py

@@ -0,0 +1,30 @@
+import json
+
+filepath = "my_data.json"
+
+def write_data(filepath, data, debug=False):
+  try:
+    with open(filepath, 'w') as f:
+      json.dump(data, f, indent=4)  # Use indent for pretty formatting
+    if debug:
+        print(f"Data written to '{filepath}'")
+  except Exception as e:
+    print(f"Error writing to file: {e}")
+
+
+def read_data(filepath, debug=False):
+  try:
+    with open(filepath, 'r') as f:
+      data = json.load(f)
+    if debug:
+        print(f"Data read from '{filepath}'")
+    return data
+  except FileNotFoundError:
+    print(f"File not found: {filepath}")
+    return None
+  except json.JSONDecodeError:
+    print(f"Error decoding JSON from file: {filepath}")
+    return None
+  except Exception as e:
+    print(f"Error reading from file: {e}")
+    return None

generator.py

@@ -1,18 +1,28 @@
+import torch
 import torch.nn as nn
 
 def conv_block(in_channels, out_channels, kernel_size=3, dilation=1):
     return nn.Sequential(
-        nn.Conv1d(in_channels, out_channels, kernel_size=kernel_size, dilation=dilation, padding=(kernel_size // 2) * dilation),
-        nn.BatchNorm1d(out_channels),
+        nn.Conv1d(
+            in_channels,
+            out_channels,
+            kernel_size=kernel_size,
+            dilation=dilation,
+            padding=(kernel_size // 2) * dilation
+        ),
+        nn.InstanceNorm1d(out_channels),
         nn.PReLU()
     )
 
 class AttentionBlock(nn.Module):
+    """
+    Simple Channel Attention Block. Learns to weight channels based on their importance.
+    """
     def __init__(self, channels):
         super(AttentionBlock, self).__init__()
         self.attention = nn.Sequential(
             nn.Conv1d(channels, channels // 4, kernel_size=1),
-            nn.ReLU(),
+            nn.ReLU(inplace=True),
             nn.Conv1d(channels // 4, channels, kernel_size=1),
             nn.Sigmoid()
         )
@@ -24,7 +34,11 @@ class AttentionBlock(nn.Module):
 class ResidualInResidualBlock(nn.Module):
     def __init__(self, channels, num_convs=3):
         super(ResidualInResidualBlock, self).__init__()
-        self.conv_layers = nn.Sequential(*[conv_block(channels, channels) for _ in range(num_convs)])
+
+        self.conv_layers = nn.Sequential(
+            *[conv_block(channels, channels) for _ in range(num_convs)]
+        )
+
         self.attention = AttentionBlock(channels)
 
     def forward(self, x):
@@ -34,19 +48,27 @@ class ResidualInResidualBlock(nn.Module):
         return x + residual
 
 class SISUGenerator(nn.Module):
-    def __init__(self, layer=4, num_rirb=4): #increased base layer and rirb amounts
+    def __init__(self, channels=16, num_rirb=4, alpha=1.0):
         super(SISUGenerator, self).__init__()
+        self.alpha = alpha
+
         self.conv1 = nn.Sequential(
-            nn.Conv1d(1, layer, kernel_size=7, padding=3),
-            nn.BatchNorm1d(layer),
+            nn.Conv1d(1, channels, kernel_size=7, padding=3),
+            nn.InstanceNorm1d(channels),
             nn.PReLU(),
         )
-        self.rir_blocks = nn.Sequential(*[ResidualInResidualBlock(layer) for _ in range(num_rirb)])
-        self.final_layer = nn.Conv1d(layer, 1, kernel_size=3, padding=1)
+
+        self.rir_blocks = nn.Sequential(
+            *[ResidualInResidualBlock(channels) for _ in range(num_rirb)]
+        )
+
+        self.final_layer = nn.Conv1d(channels, 1, kernel_size=3, padding=1)
 
     def forward(self, x):
-        residual = x
+        residual_input = x
         x = self.conv1(x)
-        x = self.rir_blocks(x)
-        x = self.final_layer(x)
-        return x + residual
+        x_rirb_out = self.rir_blocks(x)
+        learned_residual = self.final_layer(x_rirb_out)
+        output = residual_input + self.alpha * learned_residual
+
+        return output

requirements.txt

@@ -5,10 +5,8 @@ MarkupSafe==2.1.5
 mpmath==1.3.0
 networkx==3.4.2
 numpy==2.2.3
-pytorch-triton-rocm==3.2.0+git4b3bb1f8
+pillow==11.0.0
 setuptools==70.2.0
 sympy==1.13.3
-torch==2.7.0.dev20250226+rocm6.3
-torchaudio==2.6.0.dev20250226+rocm6.3
 tqdm==4.67.1
 typing_extensions==4.12.2

training.py

@@ -20,6 +20,9 @@ from data import AudioDataset
 from generator import SISUGenerator
 from discriminator import SISUDiscriminator
 
+from training_utils import discriminator_train, generator_train
+import file_utils as Data
+
 import torchaudio.transforms as T
 
 # Init script argument parser
@@ -31,95 +34,87 @@ parser.add_argument("--discriminator", type=str, default=None,
 parser.add_argument("--device", type=str, default="cpu",  help="Select device")
 parser.add_argument("--epoch", type=int, default=0, help="Current epoch for model versioning")
 parser.add_argument("--debug", action="store_true",  help="Print debug logs")
+parser.add_argument("--continue_training", action="store_true", help="Continue training using temp_generator and temp_discriminator models")
 
 args = parser.parse_args()
 
 device = torch.device(args.device if torch.cuda.is_available() else "cpu")
 print(f"Using device: {device}")
 
+# Parameters
+sample_rate = 44100
+n_fft = 1024
+win_length = n_fft
+hop_length = n_fft // 4
+n_mels = 40
+n_mfcc = 13
+
 mfcc_transform = T.MFCC(
-    sample_rate=44100,
-    n_mfcc=20,
-    melkwargs={'n_fft': 2048, 'hop_length': 256}
+    sample_rate=sample_rate,
+    n_mfcc=n_mfcc,
+    melkwargs={
+        'n_fft': n_fft,
+        'hop_length': hop_length,
+        'win_length': win_length,
+        'n_mels': n_mels,
+        'power': 1.0,
+    }
 ).to(device)
 
-def gpu_mfcc_loss(y_true, y_pred):
-    mfccs_true = mfcc_transform(y_true)
-    mfccs_pred = mfcc_transform(y_pred)
-    min_len = min(mfccs_true.shape[2], mfccs_pred.shape[2])
-    mfccs_true = mfccs_true[:, :, :min_len]
-    mfccs_pred = mfccs_pred[:, :, :min_len]
-    loss = torch.mean((mfccs_true - mfccs_pred)**2)
-    return loss
-
-def discriminator_train(high_quality, low_quality, real_labels, fake_labels):
-    optimizer_d.zero_grad()
-
-    # Forward pass for real samples
-    discriminator_decision_from_real = discriminator(high_quality[0])
-    d_loss_real = criterion_d(discriminator_decision_from_real, real_labels)
-
-    # Forward pass for fake samples (from generator output)
-    generator_output = generator(low_quality[0])
-    discriminator_decision_from_fake = discriminator(generator_output.detach())
-    d_loss_fake = criterion_d(discriminator_decision_from_fake, fake_labels)
-
-    # Combine real and fake losses
-    d_loss = (d_loss_real + d_loss_fake) / 2.0
-
-    # Backward pass and optimization
-    d_loss.backward()
-    nn.utils.clip_grad_norm_(discriminator.parameters(), max_norm=1.0)  # Gradient Clipping
-    optimizer_d.step()
-
-    return d_loss
-
-def generator_train(low_quality, high_quality, real_labels):
-    optimizer_g.zero_grad()
-
-    # Forward pass for fake samples (from generator output)
-    generator_output = generator(low_quality[0])
-
-    #mfcc_l = gpu_mfcc_loss(high_quality[0], generator_output)
-
-    discriminator_decision = discriminator(generator_output)
-    adversarial_loss = criterion_g(discriminator_decision, real_labels)
-
-    #combined_loss = adversarial_loss + 0.5 * mfcc_l
-
-    adversarial_loss.backward()
-    optimizer_g.step()
-
-    #return (generator_output, combined_loss, adversarial_loss, mfcc_l)
-    return (generator_output, adversarial_loss)
+mel_transform = T.MelSpectrogram(
+    sample_rate=sample_rate,
+    n_fft=n_fft,
+    hop_length=hop_length,
+    win_length=win_length,
+    n_mels=n_mels,
+    power=1.0  # Magnitude Mel
+).to(device)
 
+stft_transform = T.Spectrogram(
+    n_fft=n_fft,
+    win_length=win_length,
+    hop_length=hop_length
+).to(device)
 debug = args.debug
 
 # Initialize dataset and dataloader
 dataset_dir = './dataset/good'
 dataset = AudioDataset(dataset_dir, device)
+models_dir = "./models"
+os.makedirs(models_dir, exist_ok=True)
+audio_output_dir = "./output"
+os.makedirs(audio_output_dir, exist_ok=True)
 
 # ========= SINGLE =========
 
-train_data_loader = DataLoader(dataset, batch_size=256, shuffle=True)
+train_data_loader = DataLoader(dataset, batch_size=2048, shuffle=True, num_workers=24)
+
+
+# ========= MODELS =========
 
-# Initialize models and move them to device
 generator = SISUGenerator()
 discriminator = SISUDiscriminator()
 
 epoch: int = args.epoch
 
+if args.continue_training:
+    if args.generator is not None:
+        generator.load_state_dict(torch.load(args.generator, map_location=device, weights_only=True))
+    elif args.discriminator is not None:
+        discriminator.load_state_dict(torch.load(args.discriminator, map_location=device, weights_only=True))
+    else:
+        generator.load_state_dict(torch.load(f"{models_dir}/temp_generator.pt", map_location=device, weights_only=True))
+        discriminator.load_state_dict(torch.load(f"{models_dir}/temp_discriminator.pt", map_location=device, weights_only=True))
+
+        epoch_from_file = Data.read_data(f"{models_dir}/epoch_data.json")
+        epoch = epoch_from_file["epoch"] + 1
+
 generator = generator.to(device)
 discriminator = discriminator.to(device)
 
-if args.generator is not None:
-    generator.load_state_dict(torch.load(args.generator, map_location=device, weights_only=True))
-if args.discriminator is not None:
-    discriminator.load_state_dict(torch.load(args.discriminator, map_location=device, weights_only=True))
-
 # Loss
-criterion_g = nn.MSELoss()
-criterion_d = nn.BCELoss()
+criterion_g = nn.BCEWithLogitsLoss()
+criterion_d = nn.BCEWithLogitsLoss()
 
 # Optimizers
 optimizer_g = optim.Adam(generator.parameters(), lr=0.0001, betas=(0.5, 0.999))
@@ -129,61 +124,78 @@ optimizer_d = optim.Adam(discriminator.parameters(), lr=0.0001, betas=(0.5, 0.99
 scheduler_g = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_g, mode='min', factor=0.5, patience=5)
 scheduler_d = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_d, mode='min', factor=0.5, patience=5)
 
-models_dir = "models"
-os.makedirs(models_dir, exist_ok=True)
-
 def start_training():
     generator_epochs = 5000
     for generator_epoch in range(generator_epochs):
-        low_quality_audio = (torch.empty((1)), 1)
-        high_quality_audio = (torch.empty((1)), 1)
-        ai_enhanced_audio = (torch.empty((1)), 1)
+        high_quality_audio = ([torch.empty((1))], 1)
+        low_quality_audio = ([torch.empty((1))], 1)
+        ai_enhanced_audio = ([torch.empty((1))], 1)
 
         times_correct = 0
 
         # ========= TRAINING =========
-        for high_quality_clip, low_quality_clip in tqdm.tqdm(train_data_loader, desc=f"Training epoch {generator_epoch+1}/{generator_epochs}, Current epoch {epoch+1}"):
-        # for high_quality_clip, low_quality_clip in train_data_loader:
-            high_quality_sample = (high_quality_clip[0], high_quality_clip[1])
-            low_quality_sample = (low_quality_clip[0], low_quality_clip[1])
+        for training_data in tqdm.tqdm(train_data_loader, desc=f"Training epoch {generator_epoch+1}/{generator_epochs}, Current epoch {epoch+1}"):
+            ## Data structure:
+            # [[[float..., float..., float...], [float..., float..., float...]], [original_sample_rate, mangled_sample_rate]]
 
             # ========= LABELS =========
-            batch_size = high_quality_clip[0].size(0)
+            good_quality_data = training_data[0][0].to(device)
+            bad_quality_data = training_data[0][1].to(device)
+            original_sample_rate = training_data[1][0]
+            mangled_sample_rate = training_data[1][1]
+
+            batch_size = good_quality_data.size(0)
             real_labels = torch.ones(batch_size, 1).to(device)
             fake_labels = torch.zeros(batch_size, 1).to(device)
 
+            high_quality_audio = (good_quality_data, original_sample_rate)
+            low_quality_audio = (bad_quality_data, mangled_sample_rate)
+
             # ========= DISCRIMINATOR =========
             discriminator.train()
-            d_loss = discriminator_train(high_quality_sample, low_quality_sample, real_labels, fake_labels)
+            d_loss = discriminator_train(
+                good_quality_data,
+                bad_quality_data,
+                real_labels,
+                fake_labels,
+                discriminator,
+                generator,
+                criterion_d,
+                optimizer_d
+            )
 
             # ========= GENERATOR =========
             generator.train()
-            #generator_output, combined_loss, adversarial_loss, mfcc_l = generator_train(low_quality_sample, high_quality_sample, real_labels)
-            generator_output, adversarial_loss = generator_train(low_quality_sample, high_quality_sample, real_labels)
+            generator_output, combined_loss, adversarial_loss, mel_l1_tensor, log_stft_l1_tensor, mfcc_l_tensor = generator_train(
+                bad_quality_data,
+                good_quality_data,
+                real_labels,
+                generator,
+                discriminator,
+                criterion_d,
+                optimizer_g,
+                device,
+                mel_transform,
+                stft_transform,
+                mfcc_transform
+            )
 
             if debug:
-                print(d_loss, adversarial_loss)
-            scheduler_d.step(d_loss)
-            scheduler_g.step(adversarial_loss)
+                print(f"D_LOSS: {d_loss.item():.4f}, COMBINED_LOSS: {combined_loss.item():.4f}, ADVERSARIAL_LOSS: {adversarial_loss.item():.4f}, MEL_L1_LOSS: {mel_l1_tensor.item():.4f}, LOG_STFT_L1_LOSS: {log_stft_l1_tensor.item():.4f}, MFCC_LOSS: {mfcc_l_tensor.item():.4f}")
+            scheduler_d.step(d_loss.detach())
+            scheduler_g.step(adversarial_loss.detach())
 
             # ========= SAVE LATEST AUDIO =========
-            high_quality_audio = (high_quality_clip[0][0], high_quality_clip[1][0])
-            low_quality_audio = (low_quality_clip[0][0], low_quality_clip[1][0])
-            ai_enhanced_audio = (generator_output[0], high_quality_clip[1][0])
+            high_quality_audio = (good_quality_data, original_sample_rate)
+            low_quality_audio = (bad_quality_data, original_sample_rate)
+            ai_enhanced_audio = (generator_output, original_sample_rate)
+
+        torch.save(discriminator.state_dict(), f"{models_dir}/temp_discriminator.pt")
+        torch.save(generator.state_dict(), f"{models_dir}/temp_generator.pt")
 
         new_epoch = generator_epoch+epoch
+        Data.write_data(f"{models_dir}/epoch_data.json", {"epoch": new_epoch})
 
-        if generator_epoch % 10 == 0:
-            print(f"Saved epoch {new_epoch}!")
-            torchaudio.save(f"./output/epoch-{new_epoch}-audio-crap.wav", low_quality_audio[0].cpu(), high_quality_audio[1]) # <-- Because audio clip was resampled in data.py from original to crap and to original again.
-            torchaudio.save(f"./output/epoch-{new_epoch}-audio-ai.wav", ai_enhanced_audio[0].cpu(), ai_enhanced_audio[1])
-            torchaudio.save(f"./output/epoch-{new_epoch}-audio-orig.wav", high_quality_audio[0].cpu(), high_quality_audio[1])
-
-        if debug:
-            print(generator.state_dict().keys())
-            print(discriminator.state_dict().keys())
-        torch.save(discriminator.state_dict(), f"{models_dir}/discriminator_epoch_{new_epoch}.pt")
-        torch.save(generator.state_dict(), f"{models_dir}/generator_epoch_{new_epoch}.pt")
 
     torch.save(discriminator, "models/epoch-5000-discriminator.pt")
     torch.save(generator, "models/epoch-5000-generator.pt")

training_utils.py

@@ -0,0 +1,142 @@
+import torch
+import torch.nn as nn
+import torch.optim as optim
+
+import torchaudio
+import torchaudio.transforms as T
+
+def gpu_mfcc_loss(mfcc_transform, y_true, y_pred):
+    mfccs_true = mfcc_transform(y_true)
+    mfccs_pred = mfcc_transform(y_pred)
+
+    min_len = min(mfccs_true.shape[2], mfccs_pred.shape[2])
+    mfccs_true = mfccs_true[:, :, :min_len]
+    mfccs_pred = mfccs_pred[:, :, :min_len]
+
+    loss = torch.mean((mfccs_true - mfccs_pred)**2)
+    return loss
+
+def mel_spectrogram_l1_loss(mel_transform: T.MelSpectrogram, y_true: torch.Tensor, y_pred: torch.Tensor) -> torch.Tensor:
+    mel_spec_true = mel_transform(y_true)
+    mel_spec_pred = mel_transform(y_pred)
+
+    min_len = min(mel_spec_true.shape[-1], mel_spec_pred.shape[-1])
+    mel_spec_true = mel_spec_true[..., :min_len]
+    mel_spec_pred = mel_spec_pred[..., :min_len]
+
+    loss = torch.mean(torch.abs(mel_spec_true - mel_spec_pred))
+    return loss
+
+def mel_spectrogram_l2_loss(mel_transform: T.MelSpectrogram, y_true: torch.Tensor, y_pred: torch.Tensor) -> torch.Tensor:
+    mel_spec_true = mel_transform(y_true)
+    mel_spec_pred = mel_transform(y_pred)
+
+    min_len = min(mel_spec_true.shape[-1], mel_spec_pred.shape[-1])
+    mel_spec_true = mel_spec_true[..., :min_len]
+    mel_spec_pred = mel_spec_pred[..., :min_len]
+
+    loss = torch.mean((mel_spec_true - mel_spec_pred)**2)
+    return loss
+
+def log_stft_magnitude_loss(stft_transform: T.Spectrogram, y_true: torch.Tensor, y_pred: torch.Tensor, eps: float = 1e-7) -> torch.Tensor:
+    stft_mag_true = stft_transform(y_true)
+    stft_mag_pred = stft_transform(y_pred)
+
+    min_len = min(stft_mag_true.shape[-1], stft_mag_pred.shape[-1])
+    stft_mag_true = stft_mag_true[..., :min_len]
+    stft_mag_pred = stft_mag_pred[..., :min_len]
+
+    loss = torch.mean(torch.abs(torch.log(stft_mag_true + eps) - torch.log(stft_mag_pred + eps)))
+    return loss
+
+def spectral_convergence_loss(stft_transform: T.Spectrogram, y_true: torch.Tensor, y_pred: torch.Tensor, eps: float = 1e-7) -> torch.Tensor:
+    stft_mag_true = stft_transform(y_true)
+    stft_mag_pred = stft_transform(y_pred)
+
+    min_len = min(stft_mag_true.shape[-1], stft_mag_pred.shape[-1])
+    stft_mag_true = stft_mag_true[..., :min_len]
+    stft_mag_pred = stft_mag_pred[..., :min_len]
+
+    norm_true = torch.linalg.norm(stft_mag_true, ord='fro', dim=(-2, -1))
+    norm_diff = torch.linalg.norm(stft_mag_true - stft_mag_pred, ord='fro', dim=(-2, -1))
+
+    loss = torch.mean(norm_diff / (norm_true + eps))
+    return loss
+
+def discriminator_train(high_quality, low_quality, real_labels, fake_labels, discriminator, generator, criterion, optimizer):
+    optimizer.zero_grad()
+
+    # Forward pass for real samples
+    discriminator_decision_from_real = discriminator(high_quality)
+    d_loss_real = criterion(discriminator_decision_from_real, real_labels)
+
+    with torch.no_grad():
+        generator_output = generator(low_quality)
+    discriminator_decision_from_fake = discriminator(generator_output)
+    d_loss_fake = criterion(discriminator_decision_from_fake, fake_labels.expand_as(discriminator_decision_from_fake))
+
+    d_loss = (d_loss_real + d_loss_fake) / 2.0
+
+    d_loss.backward()
+    # Optional: Gradient Clipping (can be helpful)
+    # nn.utils.clip_grad_norm_(discriminator.parameters(), max_norm=1.0)  # Gradient Clipping
+    optimizer.step()
+
+    return d_loss
+
+def generator_train(
+    low_quality,
+    high_quality,
+    real_labels,
+    generator,
+    discriminator,
+    adv_criterion,
+    g_optimizer,
+    device,
+    mel_transform: T.MelSpectrogram,
+    stft_transform: T.Spectrogram,
+    mfcc_transform: T.MFCC,
+    lambda_adv: float = 1.0,
+    lambda_mel_l1: float = 10.0,
+    lambda_log_stft: float = 1.0,
+    lambda_mfcc: float = 1.0
+):
+    g_optimizer.zero_grad()
+
+    generator_output = generator(low_quality)
+
+    discriminator_decision = discriminator(generator_output)
+    adversarial_loss = adv_criterion(discriminator_decision, real_labels.expand_as(discriminator_decision))
+
+    mel_l1 = 0.0
+    log_stft_l1 = 0.0
+    mfcc_l = 0.0
+
+    # Calculate Mel L1 Loss if weight is positive
+    if lambda_mel_l1 > 0:
+        mel_l1 = mel_spectrogram_l1_loss(mel_transform, high_quality, generator_output)
+
+    # Calculate Log STFT L1 Loss if weight is positive
+    if lambda_log_stft > 0:
+        log_stft_l1 = log_stft_magnitude_loss(stft_transform, high_quality, generator_output)
+
+    # Calculate MFCC Loss if weight is positive
+    if lambda_mfcc > 0:
+        mfcc_l = gpu_mfcc_loss(mfcc_transform, high_quality, generator_output)
+
+    mel_l1_tensor = torch.tensor(mel_l1, device=device) if isinstance(mel_l1, float) else mel_l1
+    log_stft_l1_tensor = torch.tensor(log_stft_l1, device=device) if isinstance(log_stft_l1, float) else log_stft_l1
+    mfcc_l_tensor = torch.tensor(mfcc_l, device=device) if isinstance(mfcc_l, float) else mfcc_l
+
+    combined_loss = (lambda_adv * adversarial_loss) + \
+                    (lambda_mel_l1 * mel_l1_tensor) + \
+                    (lambda_log_stft * log_stft_l1_tensor) + \
+                    (lambda_mfcc * mfcc_l_tensor)
+
+    combined_loss.backward()
+    # Optional: Gradient Clipping
+    # nn.utils.clip_grad_norm_(generator.parameters(), max_norm=1.0)
+    g_optimizer.step()
+
+    # 6. Return values for logging
+    return generator_output, combined_loss, adversarial_loss, mel_l1_tensor, log_stft_l1_tensor, mfcc_l_tensor